Why AI data centres use so much electricity

What happened?

AI is part of a much bigger data-centre story. The International Energy Agency, or IEA, reports that global electricity demand from data centres grew by 17% in 2025. Demand from AI-focused data centres grew faster still, by 50%, as companies added hardware for training and running increasingly capable models.

The IEA now expects total data-centre electricity consumption to rise from about 485 TWh in 2025 to about 950 TWh in 2030 in its central projection. That is not energy used by a single chatbot. It is the combined electricity used by huge numbers of servers, network equipment, storage, cooling systems and the other infrastructure that keeps them running.

The change is physical as well as digital. Advanced AI servers pack more processing into each rack, so their power density is rising quickly. The IEA says this puts pressure on equipment such as transformers and power electronics, while the rapid changes in AI workloads can make balancing the local electricity supply more difficult.

This is why data centres have become part of current energy conversations. The question is not simply whether an individual AI task uses a little or a lot of energy. It is how much electrical power many facilities draw at the same time, where they are connected, and how efficiently they move heat away from their equipment.

The simple version

A data centre is a building full of computers. The computers process information, but they are physical objects: electrons move through circuits, transistors switch, and components get warm.

Electrical power is the rate at which energy is transferred. A more powerful server can carry out more calculations each second, but it also needs a larger rate of energy supply. The useful result is information, not a place for energy to stay. Energy carried by output signals is eventually absorbed too, so nearly all of the electrical energy supplied to the equipment ultimately becomes thermal energy in the building or its wider surroundings.

That heat must be moved away. A small computer might use a fan and a metal heat sink. A large facility can use fans, pumps, liquid cooling loops, heat exchangers and refrigeration equipment. Cooling does not destroy energy; it transfers thermal energy from the chips to another place, often outside the building.

The important word is scale. A simple text query can be much less energy-intensive than an AI-generated video or a long reasoning task, and hardware efficiency is improving. But when large numbers of people and organisations use more demanding services, the total electricity demand can still rise quickly.

Worked equations

Why it matters

Data centres are a useful example of the difference between energy and power. A facility can use a large total energy over a year, but grid engineers must also cope with its power demand at a particular moment. A sudden high demand can require more generation, transmission capacity and electrical storage even when annual energy use is known.

Cooling is now a central engineering problem. As chips become more tightly packed, heat has less surface area and less time to escape. If a chip becomes too hot, its performance, reliability and energy use can all suffer. Moving heat through metal, air or liquid is therefore part of making the computation possible.

The IEA notes that AI loads can change rapidly, which is one reason batteries may be useful around data centres. A battery does not create extra energy. It can store energy at one time and release it later, helping the local system respond while generation and demand are being balanced.

There are real choices involved: improving chip efficiency, choosing locations with suitable networks and cooling options, using cleaner electricity, and designing facilities that can reduce demand at busy times. Physics does not decide those choices alone, but it explains the limits and trade-offs behind them.

Physics you already know

The first link is electrical power. In a circuit, power can be calculated using P = IV. This is the rate at which electrical energy is transferred to a component. The same relationship applies at a much larger scale when a building supplies thousands of processors.

The flow of electric current matters because current is charge per unit time. Power supplies, cables and transformers must be designed so that they can carry the required current safely. Resistance in real conductors also causes unwanted heating, which is another energy-transfer pathway inside the system.

This story is also thermal physics. A processor is not a perfect energy converter that makes heat disappear. Thermal energy moves from hotter components to cooler surroundings by conduction, convection and sometimes radiation. Cooling systems increase the rate of that transfer, then release the energy elsewhere.

An electrical engineer might design the power distribution, protection and storage systems that make a facility safe and reliable. A data engineer might design the systems that organise data and improve how computing work is scheduled. Both careers use physical constraints rather than treating software as separate from the real world.

Science ideas to understand

A Level stretch

PUE is useful, but it is not a complete measure of environmental impact. A facility can have an excellent PUE while still using a large amount of electricity because the IT load itself is large. The source of the electricity, the timing of demand, water use and the useful work achieved also matter.

At chip scale, the heat starts with electrical losses in transistors and interconnects. Energy is dissipated when charges move through non-zero resistance and when electrical fields inside devices are repeatedly charged and discharged. More operations per second usually means more opportunities for energy transfer unless the hardware is made more efficient.

A grid has to balance electrical supply and demand continuously. The IEA highlights rapid power swings from AI workloads because a local peak can stress transformers and cables even when a site has acceptable average consumption. This is a systems problem involving electrical engineering, control and energy storage.

The IEA also makes an important distinction between efficiency per task and total demand. Hardware and software can make each task cheaper in energy terms, while increasing use and more demanding tasks can still make the overall electricity demand grow. This is a useful example of why a single number is rarely enough for a physics-based conclusion.

Key words

Quick pupil questions